a little Twitter gold:
$TSLAQ, this will be my final thread
For reasons completely unrelated to Tesla, Elon, etc., I'm going to be winding down this account in the coming weeks. I'm still short, and I'll still be following $tslaq, but can no longer be an active participant
This is an autopilot thread
2/ A long while ago, someone asked everyone in $TSLAQ what their expertise is.
Mine is in big data. Really big data. Petabytes-per-month big data. At one point, I was the product manager managing a team that built one of the largest data clusters in the world.
3/ So, $TSLAQ, I want you to take the all leap of faith that I understand big data.
And I understand what big data unlocks, and how it unlocks it: "Artificial intelligence"
The rest of this thread is about what neural networks are, in layman's terms, and why $tsla is fucked
4/ Neural networks have been around for a long time. The concept is not new - the true SCIENCE of this computer science field was made in the 70s and 80s.
The problem back then was two-fold:
- Not enough data
- Not enough compute power
5/ What's changed?
First, we now have huge, huge quantities of unstructured data. Every phone has a camera with cloud-backed photo storage. Nearly every document sits on the cloud. Every phone call recorded.
And this explosion is more recent than some remember.
6/ Second, we have more compute power. Nvidia's GPUs, Google's TPUs, etc are so much more powerful than anything built in 2008. It's not even close. We've also started mastering distributed computing, where several chips can tackle the same problem at once.
7/ These innovations - more data, more power - are what unlock neural networks and other technical AI approaches.
And we're seeing their applications now:
Discovery is becoming an AI driven process in legal proceedings
Medical imaging and diagnosis
The Amazon Alexa
8/ And image recognition.
To use image recognition, and briefly explain "how neural nets work":
Think of neural nets as a series of data transformations that ultimately produce a statistical prediction for "what is this object"
9/ The AI element is that no human has to discern what those transformations processes are. The algorithm discovers them for you
In image recognition, the AI figured out that filters and edge detection are enormously helpful transformations in determining an object. Not a human.
10/ Here's the big takeaway though:
The more transformation layers, the more data. The more precision required, the more data. And it's exponential. It's explosive.
This is why AI is so limited today. And why it's a bubble. A "buzz word"
11/ So, how do autonomous cars handle this?
For all non- $TSLA companies, they bring back the human. Humans define 3 seperate systems that feed each other:
Object recognition
Object prediction
Maneuver execution
12/ Working backwards...
In order for a car to maneuver itself, it needs to know what the things around it are going to do, and how it should handle those in its maneuvers.
In order to predict what the things around it are going to do, it needs to recognize them.
13/ You need to be able to tell the difference between a bike with a person on it from a bike, or a fire hydrant. The first might start moving, the other two aren't going to move.
14/ Key point: this is a HUMAN definition, not AI one. I'll come back to this in a moment.
Humans defined:
Recognize objects, then predict what those objects will do, then drive
Already, the data and simulations required are extreme, because each layer needs very high accuracy
15/ So... Let's get back to $TSLA.
Why will $TSLA self driving fail?
First, because they're trying a brute-force neural network, which won't work.
A) They don't have enough data. See my previous threads; they aren't sufficient data from the cars
Unroll available on Thread Reader
Value Dissenter
@ValueDissenter
Been saying this forever... I can't believe no analyst has asked "how many tb of training data does $tsla receive per day?"
They don't store the video data on the car. They don't transmit the data out of the car. $tsla is not in the race for autonomous driving. They lost.
https://twitter.com/Trumpery45/status/1051251641272262656 …
Justin
@Trumpery45
Replying to @Trumpery45 and 3 others
6/ They have no on-board storage space (witness: the dash cam needs a USB stick) so they aren't collecting the gigabytes of data per mile that Waymo etc uses to train its systems
88
9:22 PM - Oct 15, 2018
Twitter Ads info and privacy
24 people are talking about this
16/
B) Even if they were getting enough data, a brute-force neural network will demand a computer way more powerful than what's on the cars today. We're talking at least 20x-30x, probably closer to 100x
17/ LIDAR matters here.
LIDAR adds immense value in two ways:
A) Distance measurement
B) Seeing around corners
Just going to focus on A. Right now, Tesla is relying on cameras to tell how far away an object is.
Those cameras aren't precise enough, but even if they were...
18/ The data is the problem:
Drop AI and just imagine the algorithm.
On Frame A, you have to measure how far away the object is.
On Frame B, you have to again measure how far the object is, and use those changes to detect/predict movement in combination with your own movement
19/ This means that, at a bare minimum, the car must have every uncompressed, raw camera frame sitting in memory AND the frames of the past few seconds.
Cameras just aren't that good. Small pixel differences between a few frames isn't enough to distinguish how far an object is.
20/ LIDAR solves this data problem. It's an easy data feed that can quickly determine object distance and trajectory (used for object prediction).
Without it, the computer engineering problem is impossible. You can't manage and process the data on the vehicle with today's tech
21/ Without LIDAR, you're stuck with raw video data from cars on bumpy roads with low-resolution cameras loaded into a huge RAM buffer.
And from that, you'll waste a huge amount of compute time just trying to determine an object's position and movement.
22/ So, that's why I know that $TSLA will never achieve an autonomous car:
They don't and won't ever have the data required to do a brute force neural net.
And the current hardware on the vehicles is no where close to sufficient to pilot the car anyway.
23/ And there are so many other things I could talk about:
- Why mapping matters
- Computer depreciation is real
- Other AI approaches that aren't neural nets
- Data labeling problems
But the core thesis is this:
$TSLA has a fundamentally flawed approach to autonomous vehicles
24/ That's ultimately why I'm short long-term, and how I joined $TSLAQ.
Because I can absolutely guarantee that $TSLA will never, ever, ever achieve a fully self-driving car
@elonmusk has driven away all the talented AI and ML engineers that told him this.
25/ And thus $TSLA is left with second-tier computer scientists trying desperately to do what they know is impossible.
It's Theranos all over again.
Thanks, $TSLAQ. I'll still be around for a few weeks, and am happy to answer questions on this topic.
It's been fun!